Multiscale Modeling And Control Of A Plasma Etch Process

Integrated Circuits(IC)is the core segment of all intelligence systems,which also contains the highest technical content among industrial systems.Plasma etch is widely implemented in the processing of IC,especially when the manufacturing scale becomes smaller.Due to the atomic scale of the current manufacturing process,little changes in the process would lead to enormous deviation.The traditional “mistake-correction”technique is unable to maintain and optimize the current processing techniques.To improve our understandings and promote the techniques,simulation models are potential approaches.In this article,we propose a three-dimensional multiscale model for the plasma etch process.Moreover,the first-principle model and data-driven model are deeply combined in our work to generate closed-loop control and optimization system.First,a three-dimensional multiscale model is developed for the inductively coupled plasma(ICP)etch equipment.Complex chemical reactions and physical phenomena are included in the macroscopic plasma chamber.The microscopic process is highly stochastic,which consists of several etch reactions and transport phenomena.Therefore,the proposed multiscale model includes: a fluid model is used to simulate the macroscopic plasma;a kinetic Monte Carlo(k MC)model is used to simulate the microscopic etch process.By utilizing the temporal/spatial discrete method,the macroscopic and the microscopic model are computed concurrently,Second,the computational complexity of the fluid model and the k MC model prevent the direct application of these models to model-based control.To reduce the requirements for computing resources,neural network(NN)and model order reduction methods are implemented in our work to identify the simulation models.With openloop simulation data,we establish NN models to identify both the macroscopic and microscopic models in order to accelerate the simulation.Moreover,empirical eigenfunctions for partial differential equations(PDEs)system are computed by using proper orthogonal decomposition(POD),which are then applied within Galerkin’s method to derive the low-order ordinary differential equations(ODEs)system.Finally,we separately design the neural network based proportional-integral(PI)control system,the recurrent neural network(RNN)based offline optimization system,and the recurrent neural network based model predictive control(MPC)system.